A method and system for automatic segmentation of a region of interest in a DR image

Abstract

The application provides an automatic segmentation method and system for a region of interest in a DR image, and relates to the technical field of image processing. The method comprises the following steps: acquiring a DR transmission image of a welding region of an electrode lug of a power battery cell, and positioning an initial region of interest of a physical edge geometric chamfer of the electrode lug and a multi-layer metal lamination projection feature; performing gray compensation correction on a radiation hardening artifact generated in the initial region of interest due to a sudden change in the thickness of the metal, to obtain an electrode lug projection image after artifact suppression; based on the electrode lug projection image after artifact suppression, extracting a gray gradient distribution feature of the physical edge of the electrode lug and a local texture response feature at the junction of the multi-layer metal laminations; and fusing the gray gradient distribution feature and the local texture response feature to construct a spatial position constrained electrode lug edge prior guide map. The application improves the segmentation accuracy and topological integrity of the boundary of the electrode lug welding region.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an automatic segmentation method and system for regions of interest in DR images. Background Technology

[0002] In the quality inspection of electrode tab welding in power battery cells, DR (radiofrequency) imaging is widely used due to its non-destructive and efficient characteristics. However, due to abrupt changes in metal thickness in the electrode area, X-rays experience varying degrees of attenuation as they pass through metals of different thicknesses, causing a shift in the energy spectrum of the X-rays received by the detector. This results in ray hardening artifacts in the projected image. These artifacts manifest as localized grayscale distortions, with false dark or bright bands appearing near the boundaries of abrupt changes in metal thickness, interfering with the true grayscale gradient distribution at the physical edge of the electrode.

[0003] Existing segmentation methods based on thresholding, region growing, or conventional deep learning are prone to misidentifying false grayscale changes caused by artifacts as real edges when faced with such artifact interference. This leads to systematic deviations in the boundary positioning of the electrode welding area, making it difficult to accurately distinguish between the real physical contour of the geometric chamfer and the false contour induced by artifacts. Consequently, the boundary accuracy of the segmentation results cannot meet the stringent requirements for spatial positioning accuracy in industrial-grade welding defect interpretation. Summary of the Invention

[0004] This invention provides an automatic segmentation method and system for regions of interest in DR images, which improves the segmentation accuracy and topological integrity of the electrode welding area boundary.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A first aspect includes an automatic segmentation method for regions of interest in DR images, the method comprising:

[0007] Step 1: Obtain DR transmission image of the electrode welding area of ​​the power battery cell, and locate the initial region of interest of the geometric chamfer of the physical edge of the electrode and the projection features of the multi-layer metal laminations;

[0008] Step 2: Perform grayscale compensation correction on the ray hardening artifacts caused by abrupt changes in metal thickness in the initial region of interest to obtain the electrode projection image after artifact suppression.

[0009] Step 3: Based on the artifact-suppressed electrode projection image, extract the gray-level gradient distribution features of the physical edge of the electrode and the local texture response features at the junction of multiple metal sheets; fuse the gray-level gradient distribution features and the local texture response features to construct a priori guidance map of the electrode edge with spatial position constraints.

[0010] Step 4: Input the prior guidance map of the electrode edge into the multi-scale feature segmentation model to obtain the initial segmentation mask of the electrode welding area;

[0011] Step 5: Extract the set of boundary pixels of the initial segmentation mask of the electrode welding area, calculate the local curvature change rate between adjacent pixels in the set of boundary pixels, and construct curvature continuity constraint conditions that characterize the smoothness of the real physical contour of the electrode geometric chamfer.

[0012] Step 6: Based on the curvature continuity constraint, identify curvature abrupt change points in the set of boundary pixels, and perform interpolation smoothing and topological connectivity repair on the mask boundaries corresponding to the curvature abrupt change points to obtain the segmentation result of the tab welding area.

[0013] Secondly, an automatic region-of-interest segmentation system for DR images includes:

[0014] The acquisition module is used to acquire DR transmission images of the electrode welding area of ​​the power battery cell, and locate the initial region of interest of the geometric chamfer of the physical edge of the electrode and the projection features of the multi-layer metal laminations.

[0015] The correction module is used to perform grayscale compensation correction on the ray hardening artifacts caused by abrupt changes in metal thickness in the initial region of interest, so as to obtain the electrode projection image after artifact suppression.

[0016] The module is used to extract the gray-level gradient distribution features of the physical edge of the electrode and the local texture response features at the junction of multiple metal sheets based on the electrode projection image after artifact suppression; and to fuse the gray-level gradient distribution features and local texture response features to construct a priori guidance map of the electrode edge with spatial position constraints.

[0017] The processing module is used to input the prior guidance map of the electrode edge into the multi-scale feature segmentation model to obtain the initial segmentation mask of the electrode welding area;

[0018] The calculation module is used to extract the set of boundary pixels of the initial segmentation mask of the electrode welding area, calculate the local curvature change rate between adjacent pixels in the set of boundary pixels, and construct curvature continuity constraints that characterize the smoothness of the real physical contour of the electrode geometric chamfer.

[0019] The recognition module is used to identify curvature abrupt change points in the set of boundary pixels based on curvature continuity constraints, and to perform interpolation smoothing and topological connectivity repair on the mask boundaries corresponding to the curvature abrupt change points to obtain the segmentation result of the tab welding area.

[0020] Thirdly, a computing device, comprising:

[0021] One or more processors;

[0022] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0023] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0024] The above-described solution of the present invention has at least the following beneficial effects:

[0025] To address the aforementioned shortcomings, this invention corrects the ray-hardening artifacts caused by abrupt changes in metal thickness within the initial region of interest by performing gray-level compensation, thereby restoring the true gray-level distribution of the physical edges of the tabs and eliminating the misleading effect of artifacts on edge detection. Simultaneously, it constructs a spatially constrained prior guide map of the tab edges by fusing gray-level gradient distribution features and local texture response features, and introduces curvature continuity constraints based on the rate of change of local curvature to interpolate and smooth the boundary of the segmentation mask and repair its topological connectivity. This ensures that the geometric chamfered contour of the tabs in the segmentation result faithfully reflects the smoothness of the true physical edges, improving the segmentation accuracy and topological integrity of the tab welding area boundary. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating an automatic region-of-interest segmentation method in DR images provided by an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of an automatic region-of-interest segmentation system in DR images provided by an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram illustrating the process of an automatic segmentation method for regions of interest in DR images provided by an embodiment of the present invention. The method performs grayscale compensation correction on the ray hardening artifacts caused by abrupt changes in metal thickness within the initial region of interest to obtain an artifact-suppressed electrode projection image. Detailed Implementation

[0029] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0030] like Figure 1 As shown, an embodiment of the present invention proposes an automatic segmentation method for regions of interest in DR images, the method comprising the following steps:

[0031] Step 1: Obtain DR transmission image of the electrode welding area of ​​the power battery cell, and locate the initial region of interest of the geometric chamfer of the physical edge of the electrode and the projection features of the multi-layer metal laminations;

[0032] Step 2: Perform grayscale compensation correction on the ray hardening artifacts caused by abrupt changes in metal thickness in the initial region of interest to obtain the electrode projection image after artifact suppression.

[0033] Step 3: Based on the artifact-suppressed electrode projection image, extract the gray-level gradient distribution features of the physical edge of the electrode and the local texture response features at the junction of multiple metal sheets; fuse the gray-level gradient distribution features and the local texture response features to construct a priori guidance map of the electrode edge with spatial position constraints.

[0034] Step 4: Input the prior guidance map of the electrode edge into the multi-scale feature segmentation model to obtain the initial segmentation mask of the electrode welding area;

[0035] Step 5: Extract the set of boundary pixels of the initial segmentation mask of the electrode welding area, calculate the local curvature change rate between adjacent pixels in the set of boundary pixels, and construct curvature continuity constraint conditions that characterize the smoothness of the real physical contour of the electrode geometric chamfer.

[0036] Step 6: Based on the curvature continuity constraint, identify curvature abrupt change points in the set of boundary pixels, and perform interpolation smoothing and topological connectivity repair on the mask boundaries corresponding to the curvature abrupt change points to obtain the segmentation result of the tab welding area.

[0037] In this embodiment of the invention, by acquiring DR transmission images of the electrode welding area of ​​the power battery cell and locating the initial region of interest, grayscale compensation correction is performed on the ray hardening artifacts caused by abrupt changes in metal thickness. The grayscale gradient distribution features and local texture response features are fused to construct a priori guidance map of the electrode edge. This map is then input into a multi-scale feature segmentation model for deep and shallow layer feature position attention weighting and edge enhancement to generate an initial segmentation mask. The set of pixel points at the mask boundary is extracted to calculate the local curvature change rate and construct curvature continuity constraints. Subsequently, interpolation smoothing and topological connectivity repair are performed on the curvature abrupt change points. Therefore, this method overcomes the technical problems of image distortion caused by ray hardening artifacts in the electrode welding area due to abrupt changes in metal thickness, difficulty in accurately extracting the physical edge and weld features of the electrode under complex multi-layer stacked structures, and the susceptibility of the initial segmentation mask boundary to noise interference, resulting in jagged burrs and topological breaks. This method effectively suppresses image artifacts and accurately represents the prior features of the electrode and weld.

[0038] In a preferred embodiment of the present invention, step 1 above may include:

[0039] Step 1.1: Extract the global grayscale projection curve from the DR transmission image along the orthogonal direction of the tab extension direction. Identify the abrupt peak of the X-ray attenuation gradient caused by the geometric chamfer of the tab's physical edge in the global grayscale projection curve. Determine the candidate boundary position range of the tab's physical edge based on the abrupt peak of the attenuation gradient. Specifically, this includes: acquiring a DR transmission image of the welding area of ​​the power battery cell tab. The DR transmission image is a two-dimensional grayscale image. In its pixel coordinate system, the X-axis corresponds to the tab extension direction, and the Y-axis corresponds to the orthogonal direction of the tab extension direction. Based on the DR transmission image, extract the global grayscale projection curve along the Y-axis direction. The extraction method of the global grayscale projection curve is as follows: sum the grayscale values ​​of all pixels in each row of the DR transmission image along the X-axis direction to obtain the projection value corresponding to that row. Use the row number as the horizontal axis and the projection value as the vertical axis to construct the global grayscale projection curve. Specifically, let the size of the DR transmission image be... OK Column, among which Indicates the row number along the Y-axis. This represents the column number along the X-axis. Let the [number]th column be... Line 1 The column's pixel grayscale value Then the first Projection value of the row Calculate using the following formula: ;

[0040] in, Indicates the row number, with a value range of 100. ; This represents the column index, with a value range of 100. ; Representing coordinates The pixel grayscale value at that location; Indicates the first The projection value of the row.

[0041] Based on the global grayscale projection curve ,calculate The first-order difference along the Y direction yields the projected difference curve. Projection difference curve Calculate using the following formula: ;

[0042] in, Indicates the first The first-order difference of the projection at the row.

[0043] In the projection difference curve In the process, the abrupt peak position of the X-ray attenuation gradient caused by the geometric chamfer of the electrode's physical edge was determined, and the abrupt peak position of the attenuation gradient corresponds to the projection difference curve. The positions of the positive and negative peaks with the largest absolute values; let the position of the positive peak be . The negative peak position is Then the peak position and negative peak position Based on the reference, a preset offset is extended on both sides of the orthogonal direction extending towards the electrode tab. Determine the candidate boundary location range ,in The range extension width is determined by the typical projection width of the tab in DR transmission images.

[0044] Step 1.2: Within the strip region defined by the candidate boundary position interval, extract local grayscale projection morphological features along the tab extension direction. Match the local grayscale projection morphological features with a preset multi-layer metal sheet projection morphological template. The multi-layer metal sheet projection morphological template represents the layered alternating grayscale response pattern formed by the copper-aluminum composite transition sheet in the DR transmission image. Specifically, this includes: based on the obtained candidate boundary position interval, extracting local grayscale projection morphological features along the X-axis direction within the strip region defined by the candidate boundary position interval. The extraction method for local grayscale projection morphological features is: for the first... For each pixel in a column, its grayscale value sequence along the Y-axis is taken to form the grayscale profile vector corresponding to that column. The grayscale profile vectors of all columns within the strip region are arranged along the X-axis to obtain the local grayscale projection morphological feature matrix. .

[0045] The extracted local grayscale projection morphological features are matched with a preset multi-layer metal sheet projection morphological template. The template characterizes the layered alternating grayscale response pattern formed by the copper-aluminum composite transition sheet in DR transmission images. This is a reference grayscale profile vector whose element values ​​exhibit a stepped change with alternating bright and dark bands, corresponding to the difference in grayscale response produced by the different attenuation coefficients of the copper and aluminum layers to X-rays.

[0046] The matching process is performed according to the following steps: For the first column grayscale profile vector ,calculate With template The normalized cross-correlation coefficient between them is used as a matching index. Calculate using the following formula: ;

[0047] in, The dimension representing the grayscale profile. ; Represents grayscale profile vector The first in One element; Represents template vector The first in One element; express middle The mean of the elements; Represents template vector middle The mean of the elements; Indicates the first The normalized cross-correlation coefficients corresponding to the columns have a range of values. .

[0048] Step 1.3: Pixel regions within the candidate boundary location interval whose matching degree with the multi-layer metal sheet projection shape template is higher than a preset threshold are designated as the tab confirmation region. The bounding rectangle of the tab confirmation region is extended by a preset width towards the geometric chamfer direction of the tab's physical edge to obtain the initial region of interest. Specifically, this includes: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Within the candidate boundary position range Higher than the preset matching threshold The pixel area was determined as the electrode confirmation area. Preset matching threshold The value range is from 0.6 to 0.9, and the determination method is as follows: in the X-axis direction, satisfying... The continuous column intervals are used as the X-axis range of the electrode confirmation region; in the Y-axis direction, the Y-axis range of the electrode confirmation region is the candidate boundary position interval. .

[0049] confirm the electrode area The bounding rectangle extends to the geometric chamfer direction of the physical edge of the tab with a preset width. The initial region of interest is obtained. The geometric chamfering direction of the electrode's physical edge is away from the center of the electrode confirmation area and points towards the outer edge of the electrode along the Y-axis. The expansion method is to translate the outer rectangle outward by a preset width along the two boundaries of this direction. The other two boundaries remain unchanged, with a preset width. The value ranges from 0.5 to 2.0 times the Y-axis span of the polarimeter confirmation region, thus obtaining the initial region of interest. .

[0050] In this embodiment of the invention, a progressive initial region of interest (ROI) localization process is constructed, which includes global orthogonal projection, gradient abrupt change peak location, strip region local projection morphology template matching, and connected component confirmation and expansion. The X-ray attenuation gradient abrupt change peak is extracted in the orthogonal direction using the global grayscale projection curve, achieving coarse localization of the geometric chamfer of the electrode's physical edge. Local grayscale projection morphology template matching is performed along the electrode's extension direction within the candidate strip region, improving the localization accuracy from the interval level to the pixel level. This dual-layer mechanism complements each other, balancing efficiency and accuracy. The comprehensive score weighted and fused the gradient amplitude and neighborhood grayscale contrast, effectively distinguishing between true attenuation peaks and false peaks. Mean normalization and normalized cross-correlation coefficients eliminate the interference of absolute grayscale level differences and overall image brightness variations on the matching results, enhancing the generalization ability across batches of images. Eight-neighbor connected component analysis and maximum connected component screening eliminate discrete mismatched pixel patches, ensuring the spatial continuity and physical integrity of the electrode confirmation area. The adaptive expansion strategy fully covers the electrode geometric chamfer transition area and the welding heat-affected zone.

[0051] like Figure 3 As shown, in a preferred embodiment of the present invention, step 2 above may include:

[0052] Step 2.1: Within the initial region of interest, extract grayscale distribution profiles row by row along the normal direction of the geometric chamfer of the electrode's physical edge. Perform grayscale gradient analysis on each grayscale distribution profile to identify non-physical grayscale distortion segments corresponding to radiation hardening artifacts caused by abrupt changes in metal thickness. Specifically, this includes: based on the obtained initial region of interest... ,exist The grayscale distribution profile is extracted line by line along the normal direction of the geometric chamfer of the inner edge of the electrode. The normal direction is perpendicular to the local tangent of the physical edge of the electrode, which approximately corresponds to the Y-axis direction in the DR transmission image.

[0053] Let the initial region of interest be defined. Include OK List of pixels, for the first Line, its grayscale distribution profile is a line containing A one-dimensional sequence of gray values, denoted as . ,in This represents the pixel position index along the normal direction (corresponding to the X-axis direction), with a value range of... ; Indicates the row number, with a value range of 100. .

[0054] For each grayscale distribution profile line Perform grayscale gradient analysis and calculate along First gradient of direction The calculation formula is as follows: ;

[0055] in, Indicates the first Line 1 The first-order gradient value of grayscale at each pixel location.

[0056] exist In the sequence, non-physical grayscale distortion segments corresponding to radiation hardening artifacts caused by abrupt changes in metal thickness were identified. The identification method was as follows: If a continuous segment in the search gradient exceeds a preset gradient threshold, and the difference between the mean gray levels on both sides of the segment is greater than a preset gray level difference threshold, and the distance from the estimated position of a pixel within the segment to the physical edge of the tab is less than a preset distance threshold, then the segment is determined to be a non-physical gray level distortion segment. .

[0057] Step 2.2: Based on the attenuation trend of gray values ​​within the non-physical gray-level distortion segment as the distance from the physical edge of the tab changes, a local gray-level distortion compensation function is constructed. This function is then used to perform gray-level compensation on each pixel within the non-physical gray-level distortion segment, resulting in the gray values ​​of each pixel in the initial region of interest after gray-level compensation. Specifically, this includes: analyzing the identified non-physical gray-level distortion segments... The decay trend of the inner gray value with the distance from the physical edge of the electrode is used to construct a local gray-level distortion compensation function.

[0058] For the Okay, let's define the non-physical grayscale distortion segment. The estimated position of each pixel from the physical edge of the tab The distance is ,extract Internal grayscale value With distance The decay trend of the change is fitted using an exponential decay model, and a local gray-level distortion compensation function is used. Construct using the following formula: ;

[0059] in, Indicates pixel position The estimated distance from the physical edge of the electrode; Represents the artifact attenuation amplitude coefficient, derived from Determining the peak value of internal grayscale distortion; The artifact attenuation rate coefficient is represented by... The exponential rate of internal grayscale decay is determined. ; Represents a sign function, when When +1 is taken, Take -1 at time; Indicates the first Line 1 Grayscale compensation amount per pixel.

[0060] The non-physical gray-level distortion segment is compensated using a local gray-level distortion compensation function. Each pixel within the array undergoes grayscale compensation individually. The compensation method involves subtracting the grayscale compensation amount from the original grayscale value to obtain the compensated grayscale value. .

[0061] Step 2.3 involves normalizing and mapping the gray values ​​of each pixel in the initial region of interest after gray-level compensation to obtain the artifact-suppressed polar projection image. Specifically, this includes: normalizing and mapping the gray values ​​of each pixel in the initial region of interest after gray-level compensation. ,Will Normalization mapping is performed to obtain the artifact-suppressed polar projection image. The normalization mapping will then affect the initial region of interest. The grayscale values ​​of all pixels within are linearly mapped to The interval is used to obtain the artifact-suppressed polar projection image, denoted as . Its size is similar to the initial region of interest. Consistency, for OK List.

[0062] In this embodiment of the invention, a progressive artifact suppression process is constructed, which includes gradient analysis of normal direction profile lines, identification of non-physical gray-level distortion segments, exponential decay fitting of gray-level distortion compensation, and normalization mapping. By extracting gray-level distribution profile lines column by column along the geometric chamfer normal direction of the electrode's physical edge and performing gradient analysis, the location of non-physical gray-level distortion segments caused by ray hardening artifacts is achieved. A compensation function is constructed based on the exponential decay trend of gray level with distance within the distortion segment, and the distortion parameters are automatically fitted in a data-driven manner, avoiding the problem of fixed empirical parameters failing under different acquisition conditions. After compensation, the gray-level dynamic range is unified by normalization mapping, eliminating gray-level shifts that may be introduced by the compensation operation, and ensuring that the artifact suppression results have consistent gray-level performance across different image batches.

[0063] like Figure 3 As shown, in a preferred embodiment of the present invention, step 3 above may include:

[0064] Step 3.1: On the artifact-suppressed tab projection image, calculate the gray-level gradient magnitude and gradient direction pixel by pixel along the normal direction of the geometric chamfer of the tab's physical edge. Accumulate and project the gradient magnitude along the tangent of the tab's physical edge to obtain the gray-level gradient distribution characteristics that characterize the distribution law of gray-level transition intensity along the edge direction at the tab's physical edge. Specifically, this includes: based on the obtained artifact-suppressed tab projection image... ,exist The grayscale gradient magnitude is calculated pixel-by-pixel along the normal direction of the geometric chamfer of the upper edge of the tab. With gradient direction .

[0065] grayscale gradient magnitude Calculate using the following formula: ;

[0066] in, express The first partial derivative along the X-axis; express The first-order partial derivative along the Y-axis.

[0067] gradient direction Calculate using the following formula: ;

[0068] in, Representing coordinates The gradient direction angle at the location, with a value range of 100°. .

[0069] gradient magnitude Accumulated projection along the tangential direction of the physical edge of the electrode yields the gray-level gradient distribution characteristics, which characterize the distribution law of gray-level transition intensity along the edge direction at the physical edge of the electrode. Assuming the tangential direction of the physical edge of the electrode approximately corresponds to the X-axis direction, the accumulated projection method along the tangential direction is as follows: for each column along the tangential direction... The cumulative gradient magnitude at that tangential location is obtained by summing the gradient magnitudes within the neighborhood of the physical edge of the pole along the normal direction. The calculation formula is as follows: ;

[0070] in, This indicates the reference position of the physical edge of the electrode along the Y-axis. Indicates the neighborhood half-width; Indicates the first The cumulative gradient magnitude at the column is the tangential distribution value of the gray-level gradient distribution characteristics.

[0071] For all columns Repeat the above cumulative projection operation to obtain the gray-level gradient distribution features. .

[0072] Step 3.2: On the artifact-suppressed electrode projection image, guided by the layered alternating grayscale response mode corresponding to the junction of multiple metal sheets, extract the local texture primitive arrangement direction and texture primitive contrast within the sub-region. Based on the local texture primitive arrangement direction and texture primitive contrast, construct the local texture response features of the texture structure at the junction of multiple metal sheets. Specifically, this includes: based on the obtained artifact-suppressed electrode projection image... ,exist Guided by the layered alternating grayscale response mode corresponding to the junction of multiple metal sheets, the local texture primitive arrangement direction and texture primitive contrast within the sub-region are extracted.

[0073] The extraction method is: to Divide into several overlapping sub-region windows, each with a size of [size missing]. Pixel, for the center located For a sub-region window, calculate the gray-level co-occurrence matrix (GLCM) within that window, and the elements of the GLCM. Indicates the direction angle and step length Under the condition of [condition], the gray value within the sub-region window and grayscale value Frequency of simultaneous occurrence.

[0074] Extracting local texture primitive contrast based on gray-level co-occurrence matrix. and the arrangement direction of local texture primitives .

[0075] Local texture primitive contrast Calculate using the following formula: ;

[0076] in, This indicates the gray level, typically 16 or 32. This represents the element values ​​of the normalized gray-level co-occurrence matrix, i.e. ; Indicates the pixel step size; Indicates the direction angle.

[0077] Local texture primitive arrangement direction Through search Direction angle that achieves minimum value Determined, that is: ;

[0078] in, This represents a preset set of orientation angles, typically taken as... ; Indicates the direction angle The local texture primitive contrast is calculated below.

[0079] For the center position of all sub-regions Perform the above extraction operation to obtain local texture response features.

[0080] Step 3.3: Taking the geometric center of the electrode confirmation region as the spatial reference origin, the gray-level gradient distribution features are mapped to the gradient field constraint term of the electrode physical edge, and the local texture response features are mapped to the texture field constraint term at the junction of multiple metal sheets. The weighted superposition result of the gradient field constraint term and the texture field constraint term in spatial position is used as the electrode edge prior guidance map with spatial position constraint. Specifically, this includes: based on the obtained electrode confirmation region... ,Sure The geometric center is used as the spatial reference origin. .

[0081] Based on the obtained gray-level gradient distribution characteristics ,Will The gradient field constraint term is mapped to the physical edge of the pole. In spatial location The formula for calculating the value at this location is: ;

[0082] in, Represents the spatial coordinates of a pixel ; Indicates the obtained first The cumulative gradient magnitude of the column; Indicates position The normal distance to the estimated position of the physical edge of the electrode; This represents the spatial diffusion width of the gradient field constraint, with a value ranging from 3 to 10 pixels; This represents the weighting coefficient of the gradient field constraint, with a value ranging from 0.5 to 1.5.

[0083] Based on the obtained local texture response features ,Will Mapped to texture field constraints at the junction of multiple metal laminations, texture field constraints In spatial location The formula for calculating the value at this location is: ;

[0084] in, Indicates position Local texture primitive contrast at the location; Indicates position The arrangement direction of local texture primitives at a given location; Indicates the reference texture direction at the junction of multiple metal laminations; This represents the weighting coefficients of the texture field constraint.

[0085] Gradient field constraint terms With texture field constraint terms The weighted superposition results in spatial location constitute the a priori guiding map of the pole ear edge of spatial location constraints, denoted as... .

[0086] In this embodiment of the invention, a progressive feature extraction and fusion process is constructed, which includes normal direction gradient calculation and tangential accumulation, local texture structure tensor analysis and texture response extraction, and adaptive weighted superposition of two-field spaces to generate a priori guiding map. By projecting the tangential accumulation of the normal direction gradient magnitude, the two-dimensional gray-level transition information is compressed into a one-dimensional distribution feature along the edge direction, achieving a compact representation of the gray-level transition intensity at the physical edge of the electrode. Using the local gradient structure tensor as a tool, the arrangement direction and contrast of texture primitives are extracted simultaneously, and the consistency with the layered alternating gray-level response mode is enhanced by directional weighting. The texture region response effectively distinguishes the junction of multiple metal sheets from irrelevant texture regions. Taking the geometric center of the electrode confirmation region as the spatial reference origin, the spatial position weighted superposition of gradient field constraints and texture field constraints is realized through a distance-adaptive weight allocation mechanism: the region closer to the geometric center is dominated by texture field constraints to capture the layered structure features of multiple metal sheets, while the edge transition region farther from the geometric center is dominated by gradient field constraints to capture the gray-scale transition features of the electrode physical edge. The two complement each other and are fused to generate an electrode edge prior guidance map that has both edge positioning accuracy and texture discrimination capability.

[0087] In a preferred embodiment of the present invention, step 4 above may include:

[0088] Step 4.1: Input the prior guidance map of the electrode edge into the multi-scale feature segmentation model. Extract multi-scale features through the feature extraction network of the multi-scale feature segmentation model, separating deep semantic features containing high-density projection features of the weld pool region, and shallow detail features containing the real physical boundary details of the electrode. Specifically, this includes: based on the obtained prior guidance map of the electrode edge... ,Will The input is fed into a multi-scale feature segmentation model, which is a convolutional neural network based on an encoder-decoder structure. The feature extraction network of the multi-scale feature segmentation model receives the prior guidance map of the tab edge. As input, multi-scale features are extracted through cascaded convolutional and pooling layers. These multi-scale features include deep semantic features extracted by deep convolutional layers. and shallow detail features extracted from shallow convolutional layers. Deep semantic features The high-density projection features of the weld pool region are included. The weld pool region appears as a high-grayscale cluster in DR transmission images. Its response in the feature space is captured by the large receptive field of the deep convolutional kernel. The multi-scale feature segmentation model is trained using supervised learning. The training dataset contains 2000 DR transmission images of power battery cells. Each image is accompanied by a ground truth mask of the electrode welding area annotated by professional engineers. The annotations cover the electrode body, geometric chamfer, and weld pool region. The loss function is a weighted sum of DiceLoss and cross-entropy Loss, with a weight ratio of 0.7. 0.3; the optimizer is Adam, and the initial learning rate is 0.001.

[0089] Shallow details Containing detailed information about the true physical boundaries of the electrodes, which appear as abrupt changes in grayscale gradients in DR transmission images, their responses in the feature space are captured by the smaller receptive fields of shallow convolutional kernels. Let the feature extraction network be in the... The feature map output by the layer is Its spatial resolution is The number of channels is ,in, , Represents the total number of network layers and deep semantic features. Taken from The layer output, i.e. Shallow details Taken from The layer output, where ,Right now .

[0090] Step 4.2: In the deep semantic feature extraction branch, the gray-level gradient distribution features are normalized to generate a spatial attention weight map. The high-density projection features of the weld pool region in the deep semantic features are then weighted by the spatial attention weight map to obtain the weighted deep output features. Specifically, this includes: based on the obtained gray-level gradient distribution features... In the deep semantic feature extraction branch, Normalization is performed to generate a spatial attention weight map. .

[0091] Normalization processing will Values ​​mapped to Intervals are used to generate spatial attention weight graphs. Spatial attention weight map Generate using the following formula: ;

[0092] in, express In all The minimum value within the range; express In all The maximum value within the range; Indicates tangential position Spatial attention weights at each location, with values ​​ranging from 1 to 2. .

[0093] Through spatial attention weight map The obtained deep semantic features The high-density projection features of the weld pool region in the middle weld are subjected to position attention weighting, and the weighting method is as follows: Extending along spatial dimensions to Same size, for Multiplication and weighting are performed channel-by-channel and spatial-position-by-spatial to obtain the weighted deep output features. The calculation formula is as follows: ;

[0094] in, This indicates that deep semantic features are in the first... individual channels, spatial locations The value at; This indicates that the weighted deep output features are at the 1st level. individual channels, spatial locations The value at; Indicates the channel number, with a value range of 100. .

[0095] Step 4.3: In the shallow detail feature extraction branch, the local texture response features are convolved to generate an edge enhancement guiding matrix. This matrix is ​​then used to enhance the pixel gradient response of the true physical boundary of the tab in the shallow detail features, resulting in enhanced shallow output features. Specifically, this includes: based on the obtained local texture response features... In the shallow detail feature extraction branch, Local texture primitive contrast and the arrangement direction of local texture primitives Perform convolution processing to generate an edge enhancement guiding matrix. Set a direction-sensitive edge-enhancing convolution kernel Its parameters are determined by It is determined that its nuclear element value is determined by Modulation, edge enhancement guiding matrix Generate using the following formula: ;

[0096] in, Represents the neighborhood window of the convolution kernel; Represents the relative offset coordinates within the convolution kernel; Indicates direction angle The corresponding edge-enhancing convolution kernel at the offset The weight value at the location; and Representing positions respectively In offset The spatial coordinates after; Indicates position Edge enhancement guidance value at the location.

[0097] Guided by edge enhancement matrix The obtained shallow detail features Edge enhancement is performed on the pixel gradient response of the real physical boundary of the middle electrode. The enhancement method is as follows: and Multiplicative modulation is performed channel-by-channel and spatial-position-by-spatial to obtain the enhanced shallow output features. The calculation formula is as follows: ;

[0098] in, The shallow detail features are represented in the first... individual channels, spatial locations The value at; This indicates that the enhanced shallow output features are in the first... individual channels, spatial locations The value at that location, Indicates the spatial location of the edge enhancement guiding matrix. The value at that location.

[0099] Step 4.4 involves upsampling the weighted deep output features to ensure spatial resolution matches that of the enhanced shallow output features. The upsampled deep output features and the enhanced shallow output features are then concatenated and fused to obtain the fused output features. These fused output features are then input into the classifier of the multi-scale feature segmentation model to obtain the initial segmentation mask for the tab welding region. Specifically, this includes: based on the obtained weighted deep output features... and the resulting enhanced shallow output features ,Will Perform upsampling to make its spatial resolution consistent with... Maintain consistency.

[0100] Upsampling is performed using bilinear interpolation. From spatial resolution Upsampling to After upsampling, the feature map of each channel is located in space. value at Depend on The interpolation values ​​for four adjacent positions are determined by the following formula: ;

[0101] in, Represents the spatial coordinates after upsampling; The weighting coefficients for bilinear interpolation are represented by... and The distance between them is determined by reciprocal normalization; and The values ​​are 0 or 1, representing the four corner points of the interpolation neighborhood.

[0102] Upsampled deep output features Compared with enhanced shallow output features Channel splicing and fusion are performed to obtain the fused output features. Channel splicing and fusion connects two feature maps along the channel dimension.

[0103] fused output features The input is fed into the classifier of the multi-scale feature segmentation model to obtain the initial segmentation mask of the electrode welding area, denoted as . The classifier is The convolutional layer is followed by a Sigmoid activation function, which outputs the probability value of each pixel belonging to the tab welding region. The calculation formula is as follows: ;

[0104] in, This represents the convolution weight matrix of the classifier; This represents the bias term of the classifier; This represents the Sigmoid function. ; Represents pixels The probability of belonging to the electrode welding area ranges from 1 to 2. .

[0105] Will Pixels with a probability value greater than 0.5 are marked as foreground pixels. The pixels marked as background are used to obtain the initial binarized segmentation mask. .

[0106] In this embodiment of the invention, a progressive welding region segmentation process is constructed, which includes multi-scale feature separation, spatial attention weighting of deep semantic branches, edge enhancement guidance of shallow detail branches, upsampling splicing fusion, and pixel-by-pixel classification. The multi-layer downsampling structure of the backbone coding network decouples the input prior guidance map into deep semantic features and shallow detail features, effectively separating the high-density projection features of the weld pool from the physical boundary details of the tabs. A spatial attention weight map is generated using grayscale gradient distribution features as guidance, and positional attention weighting is applied to the deep semantic features, focusing the network on the high-density projection region where the weld pool is located and suppressing feature responses from irrelevant background regions. Local textures are used as the basis for further processing. Using response features as input, an edge enhancement guidance matrix is ​​generated through learnable dual-layer convolution to selectively enhance the gradient response of the real physical boundary of the tab in shallow detail features, thereby improving the feature discrimination power at the edge. The spatial resolution of deep features is restored by transposed convolution upsampling, and then channel-wise concatenated and fused with the enhanced shallow features. This allows the classifier to utilize both high-level semantic discrimination information and low-level edge detail information to achieve accurate segmentation at the boundary between the foreground and background of the welding area. The dual-branch complementary fusion design effectively balances the integrity of the overall localization of the welding area with the fineness of the boundary, improving the accuracy of the initial segmentation mask of the tab welding area under complex DR transmission images.

[0107] In a preferred embodiment of the present invention, step 5 above may include:

[0108] Step 5.1: Perform boundary tracking on the initial segmentation mask of the tab welding area, extract all pixels constituting the outer contour of the connected domain of the initial segmentation mask, and arrange all pixels into an ordered set of boundary pixels in a counterclockwise direction. Specifically, this includes: based on the obtained initial segmentation mask of the tab welding area... ,right Boundary tracking is performed to extract all pixels that constitute the outer contour of the connected component of the initial segmentation mask. The boundary tracking process is as follows: from Starting from any pixel located at the boundary between the foreground and background, search for adjacent boundary pixels sequentially along the 8-neighborhood direction until returning to the starting pixel, forming a closed boundary trajectory; arrange all pixels on the closed boundary trajectory in a counterclockwise direction to form an ordered set of boundary pixels, assuming that the set of boundary pixels contains Pixels, denoted as .

[0109] Step 5.2: In the set of boundary pixels, take three consecutive boundary pixels segment by segment along the boundary traversal direction with a preset step size for the spacing between adjacent pixels. Calculate the local boundary curvature at the position of the middle pixel based on the spatial coordinates of the three consecutive boundary pixels. Calculate the first-order difference between adjacent local boundary curvatures along the boundary traversal direction to obtain the local curvature change rate, which characterizes the degree of drastic change in boundary curvature along the physical edge of the tab. Specifically, this includes: based on the obtained set of boundary pixels... ,exist Preset step size The distance between adjacent pixels is determined by taking three consecutive boundary pixels segment by segment along the boundary traversal direction. Let the three consecutive boundary pixels currently taken be... , and Their spatial coordinates are respectively , and Based on the spatial coordinates of these three points, the position of the middle pixel is calculated. Local boundary curvature at Local boundary curvature Calculated using the curvature formula of a circle fitted at three points. Let... , , The radius of the fitted circle determined by the three points is ,but When three points are collinear, , Specifically, the local boundary curvature Calculate using the following formula: ;

[0110] in, Indicates the first The local boundary curvature at each boundary pixel, with dimensions of .

[0111] By taking the first-order difference of the local boundary curvature along the boundary ergodic direction for adjacent local boundaries, we obtain the local curvature change rate, which characterizes the degree of drastic change in boundary curvature along the physical edge of the pole. : ;

[0112] in, Indicates the first Local boundary curvature at each boundary pixel; Indicates the first The rate of change of local curvature at each boundary pixel, with dimensions of .

[0113] For all By performing the above calculations, the local curvature change rate corresponding to each boundary pixel is obtained. .

[0114] Step 5.3: Using the smoothness of the actual physical contour of the electrode geometric chamfer as the constraint target, an allowable fluctuation range for the local curvature change rate is set. Local curvature change rates within the allowable fluctuation range are judged to satisfy curvature continuity, while those exceeding the allowable fluctuation range are judged to violate curvature continuity. This constructs a curvature continuity constraint condition based on whether the local curvature change rate falls within the allowable fluctuation range. Specifically, this includes: based on the obtained local curvature change rate corresponding to each boundary pixel... Using the smoothness characteristics of the actual physical contour of the electrode geometric chamfer as a constraint target, an allowable fluctuation range for the local curvature change rate is set. The smoothness characteristics of the actual physical contour of the electrode geometric chamfer are manifested as follows: on the continuous arc segment of the electrode geometric chamfer, the change of contour curvature along the arc length direction should be slow and continuous, without any abrupt changes in curvature. Therefore, the boundary curve that conforms to the smoothness characteristics should ensure that the local curvature change rate... satisfy: ,in, The curvature change rate tolerance threshold is determined by the design curvature radius of the electrode geometry chamfer and the pixel resolution of the DR transmission image.

[0115] Will satisfy The local rate of change of curvature is determined to satisfy curvature continuity, and then... The local rate of change of curvature is determined to violate curvature continuity. Therefore, a system is constructed based on whether the local rate of change of curvature falls within the allowable fluctuation range. Curvature continuity constraint for the discrimination criterion: ;

[0116] in, Indicates the first The curvature continuity determination result at each boundary pixel point is set to 1, indicating that the curvature continuity constraint is satisfied, and 0, indicating that the curvature continuity constraint is violated.

[0117] In this embodiment of the invention, a progressive boundary curvature analysis process is constructed, which includes boundary tracking to extract ordered contour pixels, calculation of local curvature at three continuous points, and differential and statistically adaptive allowable fluctuation range and curvature continuity discrimination based on curvature change rate. The Moore neighborhood tracking algorithm extracts all pixels of the outer contour of the connected domain in a counter-clockwise direction and establishes an ordered arrangement, ensuring the spatial continuity and consistency of the traversal direction of the boundary pixels. A three-point fixed-circle method based on a preset step size is used to calculate the local boundary curvature. Adjusting the step size controls the spatial scale of the curvature calculation, avoiding discretization noise introduced by pixel-by-pixel calculation. Simultaneously, adjacent... The first-order difference of curvature, as the local curvature change rate, effectively characterizes the drastic change in the degree of boundary curvature along the edge direction. Taking the smoothness characteristics of the real physical contour of the tab geometric chamfer as the constraint target, a statistical adaptive method is used to determine the allowable fluctuation range. This allows the constraint conditions to automatically adjust the range according to the curvature distribution characteristics of the segmentation mask in different images, avoiding the incompatibility of fixed thresholds between different data batches. The final output curvature continuity constraint discrimination result clearly marks the curvature anomaly position on the boundary in point-by-point Boolean form, providing an interpretable and traceable quantitative discrimination basis for edge contour abrupt change points caused by welding defects.

[0118] In a preferred embodiment of the present invention, step 6 above may include:

[0119] Step 6.1: Substitute the local curvature change rate corresponding to each pixel in the boundary pixel set into the curvature continuity constraint condition for judgment. Mark the pixels in the boundary pixel set whose local curvature change rate exceeds the allowable fluctuation range as curvature abrupt change points. Aggregate adjacent and continuous curvature abrupt change points into curvature abrupt change segments. Specifically, this includes: according to the established curvature continuity constraint condition... The local curvature change rate corresponding to each boundary pixel is obtained. Substitute one by one Perform a judgment, if ,Right now Then the pixels in the set of boundary pixels corresponding to the local rate of curvature change Marked as a point of sudden curvature change; if ,Right now Then the pixel point Instead of marking them as curvature abrupt change points, adjacent and consecutive curvature abrupt change points are aggregated into curvature abrupt change segments. The aggregation method is as follows: scan all marked results along the boundary traversal direction, and those that satisfy... Furthermore, consecutively numbered pixels are merged into a single curvature abrupt change segment. Let the th... The curvature abrupt change section is Its starting boundary pixel is The termination boundary pixel is The number of boundary pixels contained within the segment is This yields the set of all curvature abrupt change segments. .

[0120] Step 6.2: For each curvature abrupt change segment, using the boundary pixels at both ends of the segment where no curvature abrupt change occurs as anchor points, construct an interpolation curve between the anchor points along the smooth extension trend of the real physical contour of the tab geometric chamfer. Replace the original boundary pixels within the curvature abrupt change segment with the interpolated pixels on the interpolation curve to complete the interpolation smoothing of the mask boundary corresponding to the curvature abrupt change point. Specifically, this includes: based on the obtained curvature abrupt change segment... ,by Boundary pixels where no abrupt curvature change occurs at either end are used as anchor points. The anchor points are... and ,in for The previous boundary pixel, for The next boundary pixel.

[0121] At the anchor point and Between these points, an interpolation curve is constructed along the smooth extension trend of the actual physical contour of the geometric chamfer of the tab. The interpolation curve is constructed using cubic spline interpolation, with the position of the anchor point and the tangential direction as boundary conditions.

[0122] Set anchor points The coordinates are Its local tangential direction angle is Anchor point The coordinates are Its local tangential direction angle is Local tangential direction angle and The values ​​are estimated from the local orientation of the boundary curves within the neighborhood of the corresponding anchor point.

[0123] Construct a cubic polynomial interpolation curve with the following parametric equation: ;

[0124] in, Indicates the interpolation parameters at the anchor point. Place At the anchor point Place ; and These are the undetermined coefficients of the interpolation polynomial, determined by the following boundary conditions: location conditions. , , , Tangential condition , , , .

[0125] Solve the system of linear equations corresponding to the above boundary conditions to obtain all the coefficients of the interpolation polynomial.

[0126] exist Within the interval, use a uniform step size Take the interpolation point. The interpolation curve length is determined by matching the original boundary point spacing. Let the number of interpolation points be... ,Pick , ,in Substituting the values ​​into the interpolation curve equation yields the set of interpolated pixels. interpolated pixel set The interpolated pixels replace the curvature abrupt change segment. The original boundary pixel set within the region completes the curvature abrupt change segment. The corresponding mask boundary interpolation is smoothed for all curvature abrupt change sections. Each of them Perform the above interpolation smoothing operation.

[0127] Step 6.3: After interpolating and smoothing all curvature abrupt change segments, detect whether there are discontinuous breakpoints in the boundary of the initial segmentation mask after interpolation and smoothing due to the replacement of curvature abrupt change segments. If such discontinuous breakpoints exist, perform pixel-level bridging and connectivity between the boundary endpoints on both sides of the discontinuous breakpoint along the local boundary tangential direction to fill the boundary gap and form a closed and topologically connected single connected domain boundary. The internal region of the closed boundary is used as the segmentation result of the tab welding region. Specifically, this includes: based on the result after interpolating and smoothing all curvature abrupt change segments, detecting whether there are discontinuous breakpoints in the boundary of the initial segmentation mask after interpolation and smoothing due to the replacement of curvature abrupt change segments. The discontinuous breakpoint refers to the position where the spatial distance between adjacent boundary pixels exceeds a preset connectivity threshold in the boundary pixel sequence after interpolation and smoothing. The detection method is: sequentially calculate adjacent boundary pixels along the boundary traversal direction. and Euclidean distance between .like ,in This indicates a preset connectivity threshold, set to 1.5 pixel spacing, at which point a decision is made. and There are discontinuous breakpoints between them.

[0128] If the discontinuity exists, then the boundary endpoints on both sides of the discontinuity. and Pixel-level bridging connections are established between them along the tangential direction of local boundaries. The pixel-level bridging connection method is as follows: In and The Bresenham line generation algorithm is used to calculate the coordinates of intermediate pixels. Pixels at these intermediate coordinates are then used to fill boundary gaps. This bridging and connectivity operation is repeated for all detected discontinuous points to form a closed and topologically connected single-connected domain boundary. The internal region of this closed boundary is used as the segmentation result of the tab welding area, denoted as... The In the DR image, all pixels inside the closed boundary are marked as the tab welding area, and all pixels outside the closed boundary are marked as the background, thereby completing the automatic segmentation of the region of interest.

[0129] In this embodiment of the invention, a progressive boundary repair and segmentation optimization process is constructed, which includes marking and aggregating curvature abrupt change points, smoothing with cubic spline interpolation based on anchor points, detecting breakpoints, and bridging with Bressenham lines. By substituting the local curvature change rate into the curvature continuity constraint condition for binary discrimination, isolated abrupt change points are organized into curvature abrupt change segments with spatial continuity using index continuity as the aggregation criterion, thus avoiding the boundary fragmentation problem that may be caused by point-by-point processing. Cubic spline interpolation curves are constructed with the spatial position and tangential direction of the front and rear anchor points as dual constraints to ensure that the replaced boundary is continuously connected to the original smooth contour C¹ at the endpoints, eliminating abnormal contour undulations corresponding to curvature abrupt change segments and maintaining the smoothness of the real physical contour of the pole geometric chamfer. Pixel-level boundary discontinuities that may be introduced are replaced by Chebyshev distance detection interpolation, and the Bressenham line algorithm is used to bridge the shortest path in the local tangential direction. Finally, a scanline filling algorithm is used to generate a topologically closed single connected domain segmentation result, ensuring the spatial integrity and topological connectivity of the output segmentation mask.

[0130] like Figure 2 As shown, embodiments of the present invention also provide an automatic region-of-interest segmentation system for DR images, comprising:

[0131] The acquisition module is used to acquire DR transmission images of the electrode welding area of ​​the power battery cell, and locate the initial region of interest of the geometric chamfer of the physical edge of the electrode and the projection features of the multi-layer metal laminations.

[0132] The correction module is used to perform grayscale compensation correction on the ray hardening artifacts caused by abrupt changes in metal thickness in the initial region of interest, so as to obtain the electrode projection image after artifact suppression.

[0133] The module is used to extract the gray-level gradient distribution features of the physical edge of the electrode and the local texture response features at the junction of multiple metal sheets based on the electrode projection image after artifact suppression; and to fuse the gray-level gradient distribution features and local texture response features to construct a priori guidance map of the electrode edge with spatial position constraints.

[0134] The processing module is used to input the prior guidance map of the electrode edge into the multi-scale feature segmentation model to obtain the initial segmentation mask of the electrode welding area;

[0135] The calculation module is used to extract the set of boundary pixels of the initial segmentation mask of the electrode welding area, calculate the local curvature change rate between adjacent pixels in the set of boundary pixels, and construct curvature continuity constraints that characterize the smoothness of the real physical contour of the electrode geometric chamfer.

[0136] The recognition module is used to identify curvature abrupt change points in the set of boundary pixels based on curvature continuity constraints, and to perform interpolation smoothing and topological connectivity repair on the mask boundaries corresponding to the curvature abrupt change points to obtain the segmentation result of the tab welding area.

[0137] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0138] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0139] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0140] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An automatic segmentation method for regions of interest in DR images, characterized in that, The method includes: Step 1: Obtain DR transmission images of the welding area of ​​the battery cell tabs, and locate the initial region of interest (ROI) for the geometric chamfer of the physical edge of the tabs and the projection features of the multi-layer metal laminations, including: Global grayscale projection curves are extracted from DR transmission images along the orthogonal direction of the electrode extension direction. The abrupt peak position of the X-ray attenuation gradient caused by the geometric chamfer of the physical edge of the electrode is identified in the global grayscale projection curve. The candidate boundary position range of the physical edge of the electrode is determined based on the abrupt peak position of the attenuation gradient. Within the strip area defined by the candidate boundary position interval, local grayscale projection morphological features are extracted along the extension direction of the tab. The local grayscale projection morphological features are matched with a preset multi-layer metal sheet projection morphological template. The multi-layer metal sheet projection morphological template represents the layered alternating grayscale response pattern formed by the copper-aluminum composite transition sheet in the DR transmission image. The pixel region within the candidate boundary position interval that matches the multi-layer metal sheet projection shape template with a higher than a preset threshold is taken as the tab confirmation region. The outer rectangle of the tab confirmation region is extended with a preset width in the direction of the geometric chamfer of the physical edge of the tab to obtain the initial region of interest. Step 2: Perform grayscale compensation correction on the ray hardening artifacts caused by abrupt changes in metal thickness in the initial region of interest to obtain the electrode projection image after artifact suppression. Step 3: Based on the artifact-suppressed electrode projection image, extract the gray-level gradient distribution features of the physical edge of the electrode and the local texture response features at the junction of multiple metal sheets; fuse the gray-level gradient distribution features and the local texture response features to construct a priori guidance map of the electrode edge with spatial position constraints. Step 4: Input the prior guidance map of the electrode edge into the multi-scale feature segmentation model to obtain the initial segmentation mask of the electrode welding area; Step 5: Extract the set of boundary pixels of the initial segmentation mask of the electrode welding area, calculate the local curvature change rate between adjacent pixels in the set of boundary pixels, and construct curvature continuity constraint conditions that characterize the smoothness of the real physical contour of the electrode geometric chamfer. Step 6: Based on the curvature continuity constraint, identify curvature abrupt change points in the set of boundary pixels, and perform interpolation smoothing and topological connectivity repair on the mask boundaries corresponding to the curvature abrupt change points to obtain the segmentation result of the tab welding area.

2. The automatic segmentation method for regions of interest in DR images according to claim 1, characterized in that, Step 2: Perform grayscale compensation correction on the ray hardening artifacts caused by abrupt changes in metal thickness within the initial region of interest to obtain the artifact-suppressed electrode projection image, including: Within the initial region of interest, grayscale distribution profile lines are extracted row by row along the normal direction of the geometric chamfer of the physical edge of the electrode. Grayscale gradient analysis is performed on each grayscale distribution profile line to identify the non-physical grayscale distortion segments corresponding to the radiation hardening artifacts caused by abrupt changes in metal thickness. Based on the attenuation trend of gray values ​​in non-physical gray-level distortion sections as the distance from the physical edge of the tab changes, a local gray-level distortion compensation function is constructed. The local gray-level distortion compensation function is used to perform gray-level compensation on each pixel in the non-physical gray-level distortion section one by one, so as to obtain the gray values ​​of each pixel in the initial region of interest after gray-level compensation is completed. The gray values ​​of each pixel in the initial region of interest after gray-level compensation are normalized and mapped to obtain the electrode projection image after artifact suppression.

3. The automatic segmentation method for regions of interest in DR images according to claim 2, characterized in that, Step 3: Based on the artifact-suppressed projection image of the electrode, extract the gray-level gradient distribution features of the physical edge of the electrode and the local texture response features at the junction of the multi-layer metal sheets. By fusing grayscale gradient distribution features and local texture response features, a spatially constrained epipolar edge prior guidance map is constructed, including: On the pole projection image after artifact suppression, the gray-level gradient magnitude and gradient direction are calculated pixel by pixel along the normal direction of the geometric chamfer of the pole physical edge. The gradient magnitude is accumulated and projected along the tangent of the pole physical edge to obtain the gray-level gradient distribution characteristics that characterize the distribution law of gray-level transition intensity along the edge direction at the pole physical edge. On the pole projection image after artifact suppression, guided by the layered alternating gray-scale response mode corresponding to the junction of multiple metal sheets, the local texture primitive arrangement direction and texture primitive contrast in the sub-region are extracted, and the local texture response features of the texture structure at the junction of multiple metal sheets are constructed based on the local texture primitive arrangement direction and texture primitive contrast. Using the geometric center of the electrode confirmation region as the spatial reference origin, the gray-level gradient distribution features are mapped to the gradient field constraint terms of the physical edge of the electrode, and the local texture response features are mapped to the texture field constraint terms at the junction of multiple metal sheets. The weighted superposition result of the gradient field constraint terms and the texture field constraint terms in spatial position is used as the prior guidance map of the electrode edge with spatial position constraint.

4. The automatic segmentation method for regions of interest in DR images according to claim 3, characterized in that, Step 4: Input the prior guidance map of the electrode edge into the multi-scale feature segmentation model to obtain the initial segmentation mask of the electrode welding area, including: The prior guidance map of the electrode edge is input into the multi-scale feature segmentation model. Multi-scale features are extracted through the feature extraction network of the multi-scale feature segmentation model. Deep semantic features containing high-density projection features of the weld pool region and shallow detail features containing the real physical boundary details of the electrode are separated. In the deep semantic feature extraction branch, the gray-level gradient distribution features are normalized to generate a spatial attention weight map. The high-density projection features of the weld pool region in the deep semantic features are then weighted by the spatial attention weight map to obtain the weighted deep output features. In the shallow detail feature extraction branch, the local texture response features are convolved to generate an edge enhancement guiding matrix. The pixel gradient response of the real physical boundary of the tab in the shallow detail features is enhanced by the edge enhancement guiding matrix to obtain the enhanced shallow output features. The weighted deep output features are upsampled to make the spatial resolution consistent with the enhanced shallow output features. The upsampled deep output features and the enhanced shallow output features are then channel-separated and fused to obtain the fused output features. The fused output features are then input into the classifier of the multi-scale feature segmentation model to obtain the initial segmentation mask of the tab welding area.

5. The automatic segmentation method for regions of interest in DR images according to claim 4, characterized in that, Step 5: Extract the set of boundary pixels of the initial segmentation mask of the electrode welding area, calculate the local curvature change rate between adjacent pixels in the boundary pixel set, and construct curvature continuity constraints characterizing the smoothness of the real physical contour of the electrode geometric chamfer, including: Boundary tracking is performed on the initial segmentation mask of the electrode welding area to extract all pixels that constitute the outer contour of the connected domain of the initial segmentation mask, and all pixels are arranged into an ordered set of boundary pixels in a counterclockwise direction. In the set of boundary pixels, three consecutive boundary pixels are taken segment by segment along the boundary traversal direction with a preset step size of the spacing between adjacent pixels. The local boundary curvature at the position of the middle pixel is calculated based on the spatial coordinates of the three consecutive boundary pixels. The first-order difference is obtained for the adjacent local boundary curvatures along the boundary traversal direction to obtain the local curvature change rate, which characterizes the degree of change of the boundary curvature along the physical edge of the tab. Using the smoothness of the actual physical contour of the electrode geometric chamfer as the constraint target, an allowable fluctuation range of the local curvature change rate is set. The local curvature change rate within the allowable fluctuation range is judged to satisfy curvature continuity, and the local curvature change rate exceeding the allowable fluctuation range is judged to violate curvature continuity. Thus, a curvature continuity constraint condition is constructed with whether the local curvature change rate falls into the allowable fluctuation range as the criterion.

6. The automatic segmentation method for regions of interest in DR images according to claim 5, characterized in that, Step 6: Based on the curvature continuity constraint, identify curvature abrupt change points in the boundary pixel set, and perform interpolation smoothing and topological connectivity repair on the mask boundaries corresponding to the curvature abrupt change points to obtain the segmentation result of the tab welding region, including: Substitute the local curvature change rate corresponding to each pixel in the boundary pixel set into the curvature continuity constraint condition one by one for discrimination. Mark the pixels in the boundary pixel set whose local curvature change rate exceeds the allowable fluctuation range as curvature change points. Aggregate adjacent and continuous curvature change points into curvature change segments. For each curvature change segment, the boundary pixels at both ends of the curvature change segment where no curvature change occurs are used as anchor points. An interpolation curve is constructed between the anchor points along the smooth extension trend of the real physical contour of the tab geometric chamfer. The original boundary pixels in the curvature change segment are replaced by the interpolation pixels on the interpolation curve to complete the interpolation smoothing of the mask boundary corresponding to the curvature change point. After interpolating and smoothing all curvature abrupt change segments, it is detected whether there are discontinuous breakpoints in the boundary of the initial segmentation mask after interpolation and smoothing due to the replacement of curvature abrupt change segments. If there are discontinuous breakpoints, pixel-level bridging is performed between the boundary endpoints on both sides of the discontinuous breakpoint along the local boundary tangential direction to fill the boundary gap and form a closed and topologically connected single connected domain boundary. The internal region of the closed boundary is used as the segmentation result of the tab welding area.

7. An automatic region-of-interest segmentation system for DR images, the system implementing the method as described in any one of claims 1 to 6, characterized in that, include: The acquisition module is used to acquire DR transmission images of the electrode welding area of ​​the power battery cell, and locate the initial region of interest of the geometric chamfer of the physical edge of the electrode and the projection features of the multi-layer metal laminations. The correction module is used to perform grayscale compensation correction on the ray hardening artifacts caused by abrupt changes in metal thickness in the initial region of interest, so as to obtain the electrode projection image after artifact suppression. The module is used to extract the gray-level gradient distribution features of the physical edge of the electrode and the local texture response features at the junction of multiple metal sheets based on the electrode projection image after artifact suppression; and to fuse the gray-level gradient distribution features and local texture response features to construct a priori guidance map of the electrode edge with spatial position constraints. The processing module is used to input the prior guidance map of the electrode edge into the multi-scale feature segmentation model to obtain the initial segmentation mask of the electrode welding area; The calculation module is used to extract the set of boundary pixels of the initial segmentation mask of the electrode welding area, calculate the local curvature change rate between adjacent pixels in the set of boundary pixels, and construct curvature continuity constraints that characterize the smoothness of the real physical contour of the electrode geometric chamfer. The recognition module is used to identify curvature abrupt change points in the set of boundary pixels based on curvature continuity constraints, and to perform interpolation smoothing and topological connectivity repair on the mask boundaries corresponding to the curvature abrupt change points to obtain the segmentation result of the tab welding area.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images